Keep human judgment in an AI-assisted workflow
Keep a human decision owner, define what must be checked and pause before consequential action. These are workflow design suggestions, not a measured productivity guarantee. The Power of Slow Thinking offers reflection exercises through its DEPTH framework; readers should distinguish that practical proposal from an experimentally validated method.
Put the review point before the consequence
A review step is useful only if someone can still change the outcome. Identify the action that matters: sending a message, accepting a calculation, publishing a claim or committing resources. Place a named decision owner and an explicit acceptance criterion before that action.
A small workflow to try
- Write the purpose and constraints yourself.
- Ask AI for a draft or alternatives.
- Check consequential claims against suitable independent evidence.
- Record the unresolved uncertainty.
- Have the decision owner accept, revise or reject the action.
Keep the review proportional to the task. More steps are not automatically better: the check must address a failure that matters, and the reviewer needs enough expertise to recognize it.
Observe the cost as well as the output
For one recurring task, keep a short note of review time, corrections and what happened after use. A polished draft that takes extensive repair may not improve the workflow. A faster draft may still be valuable if verification is straightforward. Your own records help distinguish those cases.
Where DEPTH fits
The Power of Slow Thinking presents practices for deliberate thought and reflection on delegation. Read its sample if you want exercises and a conceptual argument. It is not a prompt collection, a clinical intervention or a validated promise of better performance.
Read a sample before choosing

The Power of Slow Thinking
Reflection exercises for a more deliberate relationship with AI-assisted work.
Publisher-owned book. Sales support its author. Disclosure.
Sample, chapter outline and limitations →Prepared with AI assistance. Updated 2026-10-24. About our recommendations.